LOW FREQUENCY SEISMIC EXTRAPOLATION IN FULL WAVEFORM INVERSION(FWI) WITH DEEP LEARNING

• Main Author: Saputra Sigalingging, Asido
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68607
• Institution: Institut Teknologi Bandung

Full Waveform Inversion (FWI) modelling is dependent on many factors, namely the initial model, source wavelet, and low frequency of seismic data. The lack of initial model and low frequency data can affect the result of FWI modelling due to cycle skipping problems. Low frequency data is one of the crucial problems that must be tackled. The loss of low-frequency data can remove the trend of geological models. To deal with that, low-frequency data is reconstructed using deep learning methods. We use a Convolutional Neural Network (CNN) algorithm to automatically extrapolate low frequency data from bandlimited Common Shot Gather (CSG) seismic data in the time domain without pre-processing steps. The bandlimited seismic data is the input in deep learning, and the algorithm predicts low-frequency seismic data as the output. The CNN model was tested and validated with various seismic synthetic data, and the result of low-frequency prediction has good accuracy with RMSE less than 1 percent. We also applied the CNN model to real marine seismic data, Sadewa Field. The result of prediction in real data also has good accuracy about 2-3 percent RMSE. After we test the CNN model in synthetic and real data, then we run FWI modelling. We used the Marmoussi velocity model to generate synthetic seismic. The low-frequency part of the seismic Marmoussi data is predicted from CNN. The result of FWI modelling has good accuracy. These results show that our approach with deep learning seems to offer a tantalizing solution to the problem of properly initializing FWI.